Hyper-Accurate 3D Terrain Reconstruction from Drone-Based Multispectral Imagery via Adaptive Gaussian Process Regression and Iterative Structure-from-Motion Refinement

**Abstract:** This paper introduces a novel approach to generating highly accurate and dense 3D terrain models from drone-based multispectral imagery, specifically targeting applications demanding sub-decimeter precision (e.g., precision agriculture, autonomous mining). Our method combines adaptive Gaussian Process Regression (GP) for initial terrain surface estimation, leveraging spectral information to improve robustness in vegetated areas, with an iterative Structure-from-Motion (SfM) refinement process. This hybrid approach overcomes limitations of traditional SfM, which struggles with featureless or textured weaknesses caused by vegetation cover, resulting in substantially improved point cloud density and elevation accuracy compared to existing methods. The system is immediately marketable for surveying and mapping companies and directly applicable across多种展开 scenarios involving close-range 3D terrain understanding.